Scholarlink

<p>Scene text recognition has shown considerable advancements in recent years. Despite the progress, two major challenges persist in existing techniques: false positives in text representations leading to inaccuracies in recognition, and the vast scale variability of scene texts making it difficult for networks to effectively learn from diverse samples. To address these challenges, this research proposes a modification to improve text recognition accuracy on datasets containing arbitrary and irregular text samples. The proposed approach Deep Neural Encoder-Decoder network with Probabilistic Sampling (DNED-PS), leverages various neural network including CNN (Convolutional Neural Network), to limit the extraction of irrelevant features and generate more accurate text representations. Additionally, the integration of a Deep Neural Encoder-Decoder network with a modified transformer facilitates accurate text sequence generation in a bidirectional workflow, further improving overall text recognition performance. Further, the proposed model consists of three sections: Adaptive Convolutional Neural Network (CNN), Localization and Resampling (LNR), and an Encoder–Decoder network with Probabilistic Sampling (DNED-PS). Experimental evaluations demonstrate that DNED-PS achieves state-of-the-art performance across three challenging irregular text datasets. On the IIIT5K dataset, DNED-PS attains an accuracy of <b>98.3%</b>, representing a <b>3.2% absolute improvement</b> over the best existing system. On the SVTP dataset, our method achieves <b>94.3% accuracy</b>, surpassing prior baselines by <b>2.0%</b>, while on the IC15 dataset it records <b>85.6% accuracy</b>, an improvement of <b>2.6%</b> compared with the strongest competitor. These results validate the effectiveness of DNED-PS in handling irregular and distorted scene text.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Probabilistic sampling and deep neural encoder-decoder network for advancing scene text recognition on irregular datasets

  • 期刊论文

  • 发表日期：

    2026年2月6日

查看全文
  • Ratnamala S. Patil,
  • Geeta Hanji

摘要

Scene text recognition has shown considerable advancements in recent years. Despite the progress, two major challenges persist in existing techniques: false positives in text representations leading to inaccuracies in recognition, and the vast scale variability of scene texts making it difficult for networks to effectively learn from diverse samples. To address these challenges, this research proposes a modification to improve text recognition accuracy on datasets containing arbitrary and irregular text samples. The proposed approach Deep Neural Encoder-Decoder network with Probabilistic Sampling (DNED-PS), leverages various neural network including CNN (Convolutional Neural Network), to limit the extraction of irrelevant features and generate more accurate text representations. Additionally, the integration of a Deep Neural Encoder-Decoder network with a modified transformer facilitates accurate text sequence generation in a bidirectional workflow, further improving overall text recognition performance. Further, the proposed model consists of three sections: Adaptive Convolutional Neural Network (CNN), Localization and Resampling (LNR), and an Encoder–Decoder network with Probabilistic Sampling (DNED-PS). Experimental evaluations demonstrate that DNED-PS achieves state-of-the-art performance across three challenging irregular text datasets. On the IIIT5K dataset, DNED-PS attains an accuracy of 98.3%, representing a 3.2% absolute improvement over the best existing system. On the SVTP dataset, our method achieves 94.3% accuracy, surpassing prior baselines by 2.0%, while on the IC15 dataset it records 85.6% accuracy, an improvement of 2.6% compared with the strongest competitor. These results validate the effectiveness of DNED-PS in handling irregular and distorted scene text.